Fundus image segmentation via hierarchical feature learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Fundus Image Segmentation (FIS) is crucial for automated ophthalmic disease diagnosis.
• Deep learning models like U-Net are common but can lose spatial details due to pooling.
• Accurate segmentation of small vessels and lesions is vital for diagnosis.

Purpose of the Study:

• To propose a novel high-resolution hierarchical network (HHNet) for improved FIS.
• To address the limitations of existing models in preserving fine spatial details.
• To enhance the accuracy and efficiency of automated ophthalmic diagnostics.

Main Methods:

• Developed a High-resolution Feature Learning (HFL) module with increasing dilation rates.
• Constructed HHNet using three HFL modules and two feature aggregation modules.
• Implemented a coarse-to-fine approach to generate detailed segmentation maps.

Main Results:

• HHNet demonstrated competitive or superior performance across fundus lesion, vessel, and optic cup segmentation tasks.
• The proposed method effectively preserves high-resolution representations and spatial details.
• Achieved favorable segmentation performance with reduced computational cost compared to existing methods.

Conclusions:

• HHNet offers significant advantages for fundus image segmentation.
• The network's ability to maintain spatial details makes it suitable for clinical applications.
• HHNet shows potential for advancing automated diagnosis in ophthalmology.